4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * Author: Isaac Huang <isaac@clusterfs.com>
36 * 2012-05-13: Liang Zhen <liang@whamcloud.com>
37 * - percpt data for service to improve smp performance
41 #define DEBUG_SUBSYSTEM S_LNET
53 static struct smoketest_rpc {
54 spinlock_t rpc_glock; /* global lock */
55 struct srpc_service *rpc_services[SRPC_SERVICE_MAX_ID + 1];
56 struct lnet_handle_eq rpc_lnet_eq; /* _the_ LNet event queue */
57 enum srpc_state rpc_state;
58 struct srpc_counters rpc_counters;
59 __u64 rpc_matchbits; /* matchbits counter */
63 srpc_serv_portal(int svc_id)
65 return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
66 SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
70 int srpc_handle_rpc(struct swi_workitem *wi);
72 void srpc_get_counters(struct srpc_counters *cnt)
74 spin_lock(&srpc_data.rpc_glock);
75 *cnt = srpc_data.rpc_counters;
76 spin_unlock(&srpc_data.rpc_glock);
79 void srpc_set_counters(const struct srpc_counters *cnt)
81 spin_lock(&srpc_data.rpc_glock);
82 srpc_data.rpc_counters = *cnt;
83 spin_unlock(&srpc_data.rpc_glock);
87 srpc_add_bulk_page(struct srpc_bulk *bk, struct page *pg, int i, int off,
90 LASSERT(off < PAGE_SIZE);
91 LASSERT(nob > 0 && nob <= PAGE_SIZE);
93 bk->bk_iovs[i].bv_offset = off;
94 bk->bk_iovs[i].bv_page = pg;
95 bk->bk_iovs[i].bv_len = nob;
100 srpc_free_bulk(struct srpc_bulk *bk)
107 for (i = 0; i < bk->bk_niov; i++) {
108 pg = bk->bk_iovs[i].bv_page;
115 LIBCFS_FREE(bk, offsetof(struct srpc_bulk, bk_iovs[bk->bk_niov]));
119 srpc_alloc_bulk(int cpt, unsigned int bulk_off, unsigned int bulk_npg,
120 unsigned int bulk_len, int sink)
122 struct srpc_bulk *bk;
125 LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
127 LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
128 offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
130 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
134 memset(bk, 0, offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
136 bk->bk_len = bulk_len;
137 bk->bk_niov = bulk_npg;
139 for (i = 0; i < bulk_npg; i++) {
143 pg = alloc_pages_node(cfs_cpt_spread_node(lnet_cpt_table(), cpt),
146 CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
151 nob = min_t(unsigned int, bulk_off + bulk_len, PAGE_SIZE) -
153 srpc_add_bulk_page(bk, pg, i, bulk_off, nob);
166 spin_lock(&srpc_data.rpc_glock);
167 id = srpc_data.rpc_matchbits++;
168 spin_unlock(&srpc_data.rpc_glock);
173 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
174 struct srpc_service_cd *scd,
175 struct srpc_buffer *buffer)
177 memset(rpc, 0, sizeof(*rpc));
178 swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
179 srpc_serv_is_framework(scd->scd_svc) ?
180 lst_sched_serial : lst_sched_test[scd->scd_cpt]);
182 rpc->srpc_ev.ev_fired = 1; /* no event expected now */
185 rpc->srpc_reqstbuf = buffer;
186 rpc->srpc_peer = buffer->buf_peer;
187 rpc->srpc_self = buffer->buf_self;
188 LNetInvalidateMDHandle(&rpc->srpc_replymdh);
192 srpc_service_fini(struct srpc_service *svc)
194 struct srpc_service_cd *scd;
195 struct srpc_server_rpc *rpc;
196 struct srpc_buffer *buf;
200 if (!svc->sv_cpt_data)
203 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
205 if (!list_empty(&scd->scd_buf_posted))
206 q = &scd->scd_buf_posted;
207 else if (!list_empty(&scd->scd_buf_blocked))
208 q = &scd->scd_buf_blocked;
212 while (!list_empty(q)) {
213 buf = list_entry(q->next, struct srpc_buffer,
215 list_del(&buf->buf_list);
216 LIBCFS_FREE(buf, sizeof(*buf));
220 LASSERT(list_empty(&scd->scd_rpc_active));
222 while (!list_empty(&scd->scd_rpc_free)) {
223 rpc = list_entry(scd->scd_rpc_free.next,
224 struct srpc_server_rpc,
226 list_del(&rpc->srpc_list);
227 LIBCFS_FREE(rpc, sizeof(*rpc));
231 cfs_percpt_free(svc->sv_cpt_data);
232 svc->sv_cpt_data = NULL;
236 srpc_service_nrpcs(struct srpc_service *svc)
238 int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
240 return srpc_serv_is_framework(svc) ?
241 max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
244 int srpc_add_buffer(struct swi_workitem *wi);
247 srpc_service_init(struct srpc_service *svc)
249 struct srpc_service_cd *scd;
250 struct srpc_server_rpc *rpc;
255 svc->sv_shuttingdown = 0;
257 svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
258 sizeof(**svc->sv_cpt_data));
259 if (!svc->sv_cpt_data)
262 svc->sv_ncpts = srpc_serv_is_framework(svc) ?
263 1 : cfs_cpt_number(lnet_cpt_table());
264 nrpcs = srpc_service_nrpcs(svc);
266 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
269 spin_lock_init(&scd->scd_lock);
270 INIT_LIST_HEAD(&scd->scd_rpc_free);
271 INIT_LIST_HEAD(&scd->scd_rpc_active);
272 INIT_LIST_HEAD(&scd->scd_buf_posted);
273 INIT_LIST_HEAD(&scd->scd_buf_blocked);
275 scd->scd_ev.ev_data = scd;
276 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
279 * NB: don't use lst_sched_serial for adding buffer,
280 * see details in srpc_service_add_buffers()
282 swi_init_workitem(&scd->scd_buf_wi, scd,
283 srpc_add_buffer, lst_sched_test[i]);
285 if (i && srpc_serv_is_framework(svc)) {
287 * NB: framework service only needs srpc_service_cd for
288 * one partition, but we allocate for all to make
289 * it easier to implement, it will waste a little
290 * memory but nobody should care about this
295 for (j = 0; j < nrpcs; j++) {
296 LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
299 srpc_service_fini(svc);
302 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
310 srpc_add_service(struct srpc_service *sv)
314 LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
316 if (srpc_service_init(sv))
319 spin_lock(&srpc_data.rpc_glock);
321 LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
323 if (srpc_data.rpc_services[id]) {
324 spin_unlock(&srpc_data.rpc_glock);
328 srpc_data.rpc_services[id] = sv;
329 spin_unlock(&srpc_data.rpc_glock);
331 CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
335 srpc_service_fini(sv);
340 srpc_remove_service(struct srpc_service *sv)
344 spin_lock(&srpc_data.rpc_glock);
346 if (srpc_data.rpc_services[id] != sv) {
347 spin_unlock(&srpc_data.rpc_glock);
351 srpc_data.rpc_services[id] = NULL;
352 spin_unlock(&srpc_data.rpc_glock);
357 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
358 int len, int options, struct lnet_process_id peer,
359 struct lnet_handle_md *mdh, struct srpc_event *ev)
363 struct lnet_handle_me meh;
365 rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
366 local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
368 CERROR("LNetMEAttach failed: %d\n", rc);
369 LASSERT(rc == -ENOMEM);
377 md.options = options;
378 md.eq_handle = srpc_data.rpc_lnet_eq;
380 rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
382 CERROR("LNetMDAttach failed: %d\n", rc);
383 LASSERT(rc == -ENOMEM);
385 rc = LNetMEUnlink(meh);
390 CDEBUG(D_NET, "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
391 libcfs_id2str(peer), portal, matchbits);
396 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
397 int options, struct lnet_process_id peer,
398 lnet_nid_t self, struct lnet_handle_md *mdh,
399 struct srpc_event *ev)
407 md.eq_handle = srpc_data.rpc_lnet_eq;
408 md.threshold = options & LNET_MD_OP_GET ? 2 : 1;
409 md.options = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
411 rc = LNetMDBind(md, LNET_UNLINK, mdh);
413 CERROR("LNetMDBind failed: %d\n", rc);
414 LASSERT(rc == -ENOMEM);
419 * this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
420 * they're only meaningful for MDs attached to an ME (i.e. passive
423 if (options & LNET_MD_OP_PUT) {
424 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
425 portal, matchbits, 0, 0);
427 LASSERT(options & LNET_MD_OP_GET);
429 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
433 CERROR("LNet%s(%s, %d, %lld) failed: %d\n",
434 options & LNET_MD_OP_PUT ? "Put" : "Get",
435 libcfs_id2str(peer), portal, matchbits, rc);
438 * The forthcoming unlink event will complete this operation
439 * with failure, so fall through and return success here.
441 rc = LNetMDUnlink(*mdh);
444 CDEBUG(D_NET, "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
445 libcfs_id2str(peer), portal, matchbits);
451 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
452 struct lnet_handle_md *mdh, struct srpc_event *ev)
454 struct lnet_process_id any = { 0 };
456 any.nid = LNET_NID_ANY;
457 any.pid = LNET_PID_ANY;
459 return srpc_post_passive_rdma(srpc_serv_portal(service),
460 local, service, buf, len,
461 LNET_MD_OP_PUT, any, mdh, ev);
465 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
466 __must_hold(&scd->scd_lock)
468 struct srpc_service *sv = scd->scd_svc;
469 struct srpc_msg *msg = &buf->buf_msg;
472 LNetInvalidateMDHandle(&buf->buf_mdh);
473 list_add(&buf->buf_list, &scd->scd_buf_posted);
474 scd->scd_buf_nposted++;
475 spin_unlock(&scd->scd_lock);
477 rc = srpc_post_passive_rqtbuf(sv->sv_id,
478 !srpc_serv_is_framework(sv),
479 msg, sizeof(*msg), &buf->buf_mdh,
483 * At this point, a RPC (new or delayed) may have arrived in
484 * msg and its event handler has been called. So we must add
485 * buf to scd_buf_posted _before_ dropping scd_lock
487 spin_lock(&scd->scd_lock);
490 if (!sv->sv_shuttingdown)
493 spin_unlock(&scd->scd_lock);
495 * srpc_shutdown_service might have tried to unlink me
496 * when my buf_mdh was still invalid
498 LNetMDUnlink(buf->buf_mdh);
499 spin_lock(&scd->scd_lock);
503 scd->scd_buf_nposted--;
504 if (sv->sv_shuttingdown)
505 return rc; /* don't allow to change scd_buf_posted */
507 list_del(&buf->buf_list);
508 spin_unlock(&scd->scd_lock);
510 LIBCFS_FREE(buf, sizeof(*buf));
512 spin_lock(&scd->scd_lock);
517 srpc_add_buffer(struct swi_workitem *wi)
519 struct srpc_service_cd *scd = wi->swi_workitem.wi_data;
520 struct srpc_buffer *buf;
524 * it's called by workitem scheduler threads, these threads
525 * should have been set CPT affinity, so buffers will be posted
526 * on CPT local list of Portal
528 spin_lock(&scd->scd_lock);
530 while (scd->scd_buf_adjust > 0 &&
531 !scd->scd_svc->sv_shuttingdown) {
532 scd->scd_buf_adjust--; /* consume it */
533 scd->scd_buf_posting++;
535 spin_unlock(&scd->scd_lock);
537 LIBCFS_ALLOC(buf, sizeof(*buf));
539 CERROR("Failed to add new buf to service: %s\n",
540 scd->scd_svc->sv_name);
541 spin_lock(&scd->scd_lock);
546 spin_lock(&scd->scd_lock);
547 if (scd->scd_svc->sv_shuttingdown) {
548 spin_unlock(&scd->scd_lock);
549 LIBCFS_FREE(buf, sizeof(*buf));
551 spin_lock(&scd->scd_lock);
556 rc = srpc_service_post_buffer(scd, buf);
558 break; /* buf has been freed inside */
560 LASSERT(scd->scd_buf_posting > 0);
561 scd->scd_buf_posting--;
562 scd->scd_buf_total++;
563 scd->scd_buf_low = max(2, scd->scd_buf_total / 4);
567 scd->scd_buf_err_stamp = ktime_get_real_seconds();
568 scd->scd_buf_err = rc;
570 LASSERT(scd->scd_buf_posting > 0);
571 scd->scd_buf_posting--;
574 spin_unlock(&scd->scd_lock);
579 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
581 struct srpc_service_cd *scd;
585 LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
587 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
588 spin_lock(&scd->scd_lock);
590 scd->scd_buf_err = 0;
591 scd->scd_buf_err_stamp = 0;
592 scd->scd_buf_posting = 0;
593 scd->scd_buf_adjust = nbuffer;
594 /* start to post buffers */
595 swi_schedule_workitem(&scd->scd_buf_wi);
596 spin_unlock(&scd->scd_lock);
598 /* framework service only post buffer for one partition */
599 if (srpc_serv_is_framework(sv))
603 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
604 spin_lock(&scd->scd_lock);
606 * NB: srpc_service_add_buffers() can be called inside
607 * thread context of lst_sched_serial, and we don't normally
608 * allow to sleep inside thread context of WI scheduler
609 * because it will block current scheduler thread from doing
610 * anything else, even worse, it could deadlock if it's
611 * waiting on result from another WI of the same scheduler.
612 * However, it's safe at here because scd_buf_wi is scheduled
613 * by thread in a different WI scheduler (lst_sched_test),
614 * so we don't have any risk of deadlock, though this could
615 * block all WIs pending on lst_sched_serial for a moment
616 * which is not good but not fatal.
618 lst_wait_until(scd->scd_buf_err ||
619 (!scd->scd_buf_adjust &&
620 !scd->scd_buf_posting),
621 scd->scd_lock, "waiting for adding buffer\n");
623 if (scd->scd_buf_err && !rc)
624 rc = scd->scd_buf_err;
626 spin_unlock(&scd->scd_lock);
633 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
635 struct srpc_service_cd *scd;
639 LASSERT(!sv->sv_shuttingdown);
641 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
642 spin_lock(&scd->scd_lock);
644 num = scd->scd_buf_total + scd->scd_buf_posting;
645 scd->scd_buf_adjust -= min(nbuffer, num);
647 spin_unlock(&scd->scd_lock);
651 /* returns 1 if sv has finished, otherwise 0 */
653 srpc_finish_service(struct srpc_service *sv)
655 struct srpc_service_cd *scd;
656 struct srpc_server_rpc *rpc;
659 LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
661 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
662 spin_lock(&scd->scd_lock);
663 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
664 spin_unlock(&scd->scd_lock);
668 if (scd->scd_buf_nposted > 0) {
669 CDEBUG(D_NET, "waiting for %d posted buffers to unlink\n",
670 scd->scd_buf_nposted);
671 spin_unlock(&scd->scd_lock);
675 if (list_empty(&scd->scd_rpc_active)) {
676 spin_unlock(&scd->scd_lock);
680 rpc = list_entry(scd->scd_rpc_active.next,
681 struct srpc_server_rpc, srpc_list);
682 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, wi %s scheduled %d running %d, ev fired %d type %d status %d lnet %d\n",
683 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
684 swi_state2str(rpc->srpc_wi.swi_state),
685 rpc->srpc_wi.swi_workitem.wi_scheduled,
686 rpc->srpc_wi.swi_workitem.wi_running,
687 rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
688 rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
689 spin_unlock(&scd->scd_lock);
693 /* no lock needed from now on */
694 srpc_service_fini(sv);
698 /* called with sv->sv_lock held */
700 srpc_service_recycle_buffer(struct srpc_service_cd *scd,
701 struct srpc_buffer *buf)
702 __must_hold(&scd->scd_lock)
704 if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
705 if (srpc_service_post_buffer(scd, buf)) {
706 CWARN("Failed to post %s buffer\n",
707 scd->scd_svc->sv_name);
712 /* service is shutting down, or we want to recycle some buffers */
713 scd->scd_buf_total--;
715 if (scd->scd_buf_adjust < 0) {
716 scd->scd_buf_adjust++;
717 if (scd->scd_buf_adjust < 0 &&
718 !scd->scd_buf_total && !scd->scd_buf_posting) {
720 "Try to recycle %d buffers but nothing left\n",
721 scd->scd_buf_adjust);
722 scd->scd_buf_adjust = 0;
726 spin_unlock(&scd->scd_lock);
727 LIBCFS_FREE(buf, sizeof(*buf));
728 spin_lock(&scd->scd_lock);
732 srpc_abort_service(struct srpc_service *sv)
734 struct srpc_service_cd *scd;
735 struct srpc_server_rpc *rpc;
738 CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
739 sv->sv_id, sv->sv_name);
741 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
742 spin_lock(&scd->scd_lock);
745 * schedule in-flight RPCs to notice the abort, NB:
746 * racing with incoming RPCs; complete fix should make test
747 * RPCs carry session ID in its headers
749 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
750 rpc->srpc_aborted = 1;
751 swi_schedule_workitem(&rpc->srpc_wi);
754 spin_unlock(&scd->scd_lock);
759 srpc_shutdown_service(struct srpc_service *sv)
761 struct srpc_service_cd *scd;
762 struct srpc_server_rpc *rpc;
763 struct srpc_buffer *buf;
766 CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
767 sv->sv_id, sv->sv_name);
769 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
770 spin_lock(&scd->scd_lock);
772 sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
774 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
775 spin_unlock(&scd->scd_lock);
777 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
778 spin_lock(&scd->scd_lock);
780 /* schedule in-flight RPCs to notice the shutdown */
781 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
782 swi_schedule_workitem(&rpc->srpc_wi);
784 spin_unlock(&scd->scd_lock);
787 * OK to traverse scd_buf_posted without lock, since no one
788 * touches scd_buf_posted now
790 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
791 LNetMDUnlink(buf->buf_mdh);
796 srpc_send_request(struct srpc_client_rpc *rpc)
798 struct srpc_event *ev = &rpc->crpc_reqstev;
803 ev->ev_type = SRPC_REQUEST_SENT;
805 rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
806 rpc->crpc_service, &rpc->crpc_reqstmsg,
807 sizeof(struct srpc_msg), LNET_MD_OP_PUT,
808 rpc->crpc_dest, LNET_NID_ANY,
809 &rpc->crpc_reqstmdh, ev);
811 LASSERT(rc == -ENOMEM);
812 ev->ev_fired = 1; /* no more event expected */
818 srpc_prepare_reply(struct srpc_client_rpc *rpc)
820 struct srpc_event *ev = &rpc->crpc_replyev;
821 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
826 ev->ev_type = SRPC_REPLY_RCVD;
828 *id = srpc_next_id();
830 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
832 sizeof(struct srpc_msg),
833 LNET_MD_OP_PUT, rpc->crpc_dest,
834 &rpc->crpc_replymdh, ev);
836 LASSERT(rc == -ENOMEM);
837 ev->ev_fired = 1; /* no more event expected */
843 srpc_prepare_bulk(struct srpc_client_rpc *rpc)
845 struct srpc_bulk *bk = &rpc->crpc_bulk;
846 struct srpc_event *ev = &rpc->crpc_bulkev;
847 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
851 LASSERT(bk->bk_niov <= LNET_MAX_IOV);
854 return 0; /* nothing to do */
856 opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
861 ev->ev_type = SRPC_BULK_REQ_RCVD;
863 *id = srpc_next_id();
865 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
866 &bk->bk_iovs[0], bk->bk_niov, opt,
867 rpc->crpc_dest, &bk->bk_mdh, ev);
869 LASSERT(rc == -ENOMEM);
870 ev->ev_fired = 1; /* no more event expected */
876 srpc_do_bulk(struct srpc_server_rpc *rpc)
878 struct srpc_event *ev = &rpc->srpc_ev;
879 struct srpc_bulk *bk = rpc->srpc_bulk;
880 __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
886 opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
891 ev->ev_type = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
893 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
894 &bk->bk_iovs[0], bk->bk_niov, opt,
895 rpc->srpc_peer, rpc->srpc_self,
898 ev->ev_fired = 1; /* no more event expected */
902 /* only called from srpc_handle_rpc */
904 srpc_server_rpc_done(struct srpc_server_rpc *rpc, int status)
906 struct srpc_service_cd *scd = rpc->srpc_scd;
907 struct srpc_service *sv = scd->scd_svc;
908 struct srpc_buffer *buffer;
910 LASSERT(status || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
912 rpc->srpc_status = status;
914 CDEBUG_LIMIT(!status ? D_NET : D_NETERROR,
915 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
916 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
917 swi_state2str(rpc->srpc_wi.swi_state), status);
920 spin_lock(&srpc_data.rpc_glock);
921 srpc_data.rpc_counters.rpcs_dropped++;
922 spin_unlock(&srpc_data.rpc_glock);
926 (*rpc->srpc_done) (rpc);
927 LASSERT(!rpc->srpc_bulk);
929 spin_lock(&scd->scd_lock);
931 if (rpc->srpc_reqstbuf) {
933 * NB might drop sv_lock in srpc_service_recycle_buffer, but
934 * sv won't go away for scd_rpc_active must not be empty
936 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
937 rpc->srpc_reqstbuf = NULL;
940 list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
943 * No one can schedule me now since:
944 * - I'm not on scd_rpc_active.
945 * - all LNet events have been fired.
946 * Cancel pending schedules and prevent future schedule attempts:
948 LASSERT(rpc->srpc_ev.ev_fired);
949 swi_exit_workitem(&rpc->srpc_wi);
951 if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
952 buffer = list_entry(scd->scd_buf_blocked.next,
953 struct srpc_buffer, buf_list);
954 list_del(&buffer->buf_list);
956 srpc_init_server_rpc(rpc, scd, buffer);
957 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
958 swi_schedule_workitem(&rpc->srpc_wi);
960 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
963 spin_unlock(&scd->scd_lock);
966 /* handles an incoming RPC */
968 srpc_handle_rpc(struct swi_workitem *wi)
970 struct srpc_server_rpc *rpc = wi->swi_workitem.wi_data;
971 struct srpc_service_cd *scd = rpc->srpc_scd;
972 struct srpc_service *sv = scd->scd_svc;
973 struct srpc_event *ev = &rpc->srpc_ev;
976 LASSERT(wi == &rpc->srpc_wi);
978 spin_lock(&scd->scd_lock);
980 if (sv->sv_shuttingdown || rpc->srpc_aborted) {
981 spin_unlock(&scd->scd_lock);
984 LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
985 LNetMDUnlink(rpc->srpc_replymdh);
987 if (ev->ev_fired) { /* no more event, OK to finish */
988 srpc_server_rpc_done(rpc, -ESHUTDOWN);
994 spin_unlock(&scd->scd_lock);
996 switch (wi->swi_state) {
999 case SWI_STATE_NEWBORN: {
1000 struct srpc_msg *msg;
1001 struct srpc_generic_reply *reply;
1003 msg = &rpc->srpc_reqstbuf->buf_msg;
1004 reply = &rpc->srpc_replymsg.msg_body.reply;
1006 if (!msg->msg_magic) {
1007 /* moaned already in srpc_lnet_ev_handler */
1008 srpc_server_rpc_done(rpc, EBADMSG);
1012 srpc_unpack_msg_hdr(msg);
1013 if (msg->msg_version != SRPC_MSG_VERSION) {
1014 CWARN("Version mismatch: %u, %u expected, from %s\n",
1015 msg->msg_version, SRPC_MSG_VERSION,
1016 libcfs_id2str(rpc->srpc_peer));
1017 reply->status = EPROTO;
1018 /* drop through and send reply */
1021 rc = (*sv->sv_handler)(rpc);
1022 LASSERT(!reply->status || !rpc->srpc_bulk);
1024 srpc_server_rpc_done(rpc, rc);
1029 wi->swi_state = SWI_STATE_BULK_STARTED;
1031 if (rpc->srpc_bulk) {
1032 rc = srpc_do_bulk(rpc);
1034 return 0; /* wait for bulk */
1036 LASSERT(ev->ev_fired);
1040 case SWI_STATE_BULK_STARTED:
1041 LASSERT(!rpc->srpc_bulk || ev->ev_fired);
1043 if (rpc->srpc_bulk) {
1046 if (sv->sv_bulk_ready)
1047 rc = (*sv->sv_bulk_ready) (rpc, rc);
1050 srpc_server_rpc_done(rpc, rc);
1055 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1056 rc = srpc_send_reply(rpc);
1058 return 0; /* wait for reply */
1059 srpc_server_rpc_done(rpc, rc);
1062 case SWI_STATE_REPLY_SUBMITTED:
1063 if (!ev->ev_fired) {
1064 CERROR("RPC %p: bulk %p, service %d\n",
1065 rpc, rpc->srpc_bulk, sv->sv_id);
1066 CERROR("Event: status %d, type %d, lnet %d\n",
1067 ev->ev_status, ev->ev_type, ev->ev_lnet);
1068 LASSERT(ev->ev_fired);
1071 wi->swi_state = SWI_STATE_DONE;
1072 srpc_server_rpc_done(rpc, ev->ev_status);
1080 srpc_client_rpc_expired(void *data)
1082 struct srpc_client_rpc *rpc = data;
1084 CWARN("Client RPC expired: service %d, peer %s, timeout %d.\n",
1085 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1088 spin_lock(&rpc->crpc_lock);
1090 rpc->crpc_timeout = 0;
1091 srpc_abort_rpc(rpc, -ETIMEDOUT);
1093 spin_unlock(&rpc->crpc_lock);
1095 spin_lock(&srpc_data.rpc_glock);
1096 srpc_data.rpc_counters.rpcs_expired++;
1097 spin_unlock(&srpc_data.rpc_glock);
1101 srpc_add_client_rpc_timer(struct srpc_client_rpc *rpc)
1103 struct stt_timer *timer = &rpc->crpc_timer;
1105 if (!rpc->crpc_timeout)
1108 INIT_LIST_HEAD(&timer->stt_list);
1109 timer->stt_data = rpc;
1110 timer->stt_func = srpc_client_rpc_expired;
1111 timer->stt_expires = ktime_get_real_seconds() + rpc->crpc_timeout;
1112 stt_add_timer(timer);
1116 * Called with rpc->crpc_lock held.
1118 * Upon exit the RPC expiry timer is not queued and the handler is not
1119 * running on any CPU.
1122 srpc_del_client_rpc_timer(struct srpc_client_rpc *rpc)
1124 /* timer not planted or already exploded */
1125 if (!rpc->crpc_timeout)
1128 /* timer successfully defused */
1129 if (stt_del_timer(&rpc->crpc_timer))
1132 /* timer detonated, wait for it to explode */
1133 while (rpc->crpc_timeout) {
1134 spin_unlock(&rpc->crpc_lock);
1138 spin_lock(&rpc->crpc_lock);
1143 srpc_client_rpc_done(struct srpc_client_rpc *rpc, int status)
1145 struct swi_workitem *wi = &rpc->crpc_wi;
1147 LASSERT(status || wi->swi_state == SWI_STATE_DONE);
1149 spin_lock(&rpc->crpc_lock);
1151 rpc->crpc_closed = 1;
1152 if (!rpc->crpc_status)
1153 rpc->crpc_status = status;
1155 srpc_del_client_rpc_timer(rpc);
1157 CDEBUG_LIMIT(!status ? D_NET : D_NETERROR,
1158 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1159 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1160 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1163 * No one can schedule me now since:
1164 * - RPC timer has been defused.
1165 * - all LNet events have been fired.
1166 * - crpc_closed has been set, preventing srpc_abort_rpc from
1168 * Cancel pending schedules and prevent future schedule attempts:
1170 LASSERT(!srpc_event_pending(rpc));
1171 swi_exit_workitem(wi);
1173 spin_unlock(&rpc->crpc_lock);
1175 (*rpc->crpc_done)(rpc);
1178 /* sends an outgoing RPC */
1180 srpc_send_rpc(struct swi_workitem *wi)
1183 struct srpc_client_rpc *rpc;
1184 struct srpc_msg *reply;
1189 rpc = wi->swi_workitem.wi_data;
1192 LASSERT(wi == &rpc->crpc_wi);
1194 reply = &rpc->crpc_replymsg;
1195 do_bulk = rpc->crpc_bulk.bk_niov > 0;
1197 spin_lock(&rpc->crpc_lock);
1199 if (rpc->crpc_aborted) {
1200 spin_unlock(&rpc->crpc_lock);
1204 spin_unlock(&rpc->crpc_lock);
1206 switch (wi->swi_state) {
1209 case SWI_STATE_NEWBORN:
1210 LASSERT(!srpc_event_pending(rpc));
1212 rc = srpc_prepare_reply(rpc);
1214 srpc_client_rpc_done(rpc, rc);
1218 rc = srpc_prepare_bulk(rpc);
1222 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1223 rc = srpc_send_request(rpc);
1226 case SWI_STATE_REQUEST_SUBMITTED:
1228 * CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1229 * order; however, they're processed in a strict order:
1230 * rqt, rpy, and bulk.
1232 if (!rpc->crpc_reqstev.ev_fired)
1235 rc = rpc->crpc_reqstev.ev_status;
1239 wi->swi_state = SWI_STATE_REQUEST_SENT;
1240 /* perhaps more events, fall thru */
1241 case SWI_STATE_REQUEST_SENT: {
1242 enum srpc_msg_type type = srpc_service2reply(rpc->crpc_service);
1244 if (!rpc->crpc_replyev.ev_fired)
1247 rc = rpc->crpc_replyev.ev_status;
1251 srpc_unpack_msg_hdr(reply);
1252 if (reply->msg_type != type ||
1253 (reply->msg_magic != SRPC_MSG_MAGIC &&
1254 reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1255 CWARN("Bad message from %s: type %u (%d expected), magic %u (%d expected).\n",
1256 libcfs_id2str(rpc->crpc_dest),
1257 reply->msg_type, type,
1258 reply->msg_magic, SRPC_MSG_MAGIC);
1263 if (do_bulk && reply->msg_body.reply.status) {
1264 CWARN("Remote error %d at %s, unlink bulk buffer in case peer didn't initiate bulk transfer\n",
1265 reply->msg_body.reply.status,
1266 libcfs_id2str(rpc->crpc_dest));
1267 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1270 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1272 case SWI_STATE_REPLY_RECEIVED:
1273 if (do_bulk && !rpc->crpc_bulkev.ev_fired)
1276 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1279 * Bulk buffer was unlinked due to remote error. Clear error
1280 * since reply buffer still contains valid data.
1281 * NB rpc->crpc_done shouldn't look into bulk data in case of
1284 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1285 !rpc->crpc_status && reply->msg_body.reply.status)
1288 wi->swi_state = SWI_STATE_DONE;
1289 srpc_client_rpc_done(rpc, rc);
1294 spin_lock(&rpc->crpc_lock);
1295 srpc_abort_rpc(rpc, rc);
1296 spin_unlock(&rpc->crpc_lock);
1300 if (rpc->crpc_aborted) {
1301 LNetMDUnlink(rpc->crpc_reqstmdh);
1302 LNetMDUnlink(rpc->crpc_replymdh);
1303 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1305 if (!srpc_event_pending(rpc)) {
1306 srpc_client_rpc_done(rpc, -EINTR);
1313 struct srpc_client_rpc *
1314 srpc_create_client_rpc(struct lnet_process_id peer, int service,
1315 int nbulkiov, int bulklen,
1316 void (*rpc_done)(struct srpc_client_rpc *),
1317 void (*rpc_fini)(struct srpc_client_rpc *), void *priv)
1319 struct srpc_client_rpc *rpc;
1321 LIBCFS_ALLOC(rpc, offsetof(struct srpc_client_rpc,
1322 crpc_bulk.bk_iovs[nbulkiov]));
1326 srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1327 bulklen, rpc_done, rpc_fini, priv);
1331 /* called with rpc->crpc_lock held */
1333 srpc_abort_rpc(struct srpc_client_rpc *rpc, int why)
1337 if (rpc->crpc_aborted || /* already aborted */
1338 rpc->crpc_closed) /* callback imminent */
1341 CDEBUG(D_NET, "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1342 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1343 swi_state2str(rpc->crpc_wi.swi_state), why);
1345 rpc->crpc_aborted = 1;
1346 rpc->crpc_status = why;
1347 swi_schedule_workitem(&rpc->crpc_wi);
1350 /* called with rpc->crpc_lock held */
1352 srpc_post_rpc(struct srpc_client_rpc *rpc)
1354 LASSERT(!rpc->crpc_aborted);
1355 LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
1357 CDEBUG(D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1358 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1361 srpc_add_client_rpc_timer(rpc);
1362 swi_schedule_workitem(&rpc->crpc_wi);
1366 srpc_send_reply(struct srpc_server_rpc *rpc)
1368 struct srpc_event *ev = &rpc->srpc_ev;
1369 struct srpc_msg *msg = &rpc->srpc_replymsg;
1370 struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1371 struct srpc_service_cd *scd = rpc->srpc_scd;
1372 struct srpc_service *sv = scd->scd_svc;
1377 rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1379 spin_lock(&scd->scd_lock);
1381 if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1383 * Repost buffer before replying since test client
1384 * might send me another RPC once it gets the reply
1386 if (srpc_service_post_buffer(scd, buffer))
1387 CWARN("Failed to repost %s buffer\n", sv->sv_name);
1388 rpc->srpc_reqstbuf = NULL;
1391 spin_unlock(&scd->scd_lock);
1395 ev->ev_type = SRPC_REPLY_SENT;
1397 msg->msg_magic = SRPC_MSG_MAGIC;
1398 msg->msg_version = SRPC_MSG_VERSION;
1399 msg->msg_type = srpc_service2reply(sv->sv_id);
1401 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1402 sizeof(*msg), LNET_MD_OP_PUT,
1403 rpc->srpc_peer, rpc->srpc_self,
1404 &rpc->srpc_replymdh, ev);
1406 ev->ev_fired = 1; /* no more event expected */
1410 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1412 srpc_lnet_ev_handler(struct lnet_event *ev)
1414 struct srpc_service_cd *scd;
1415 struct srpc_event *rpcev = ev->md.user_ptr;
1416 struct srpc_client_rpc *crpc;
1417 struct srpc_server_rpc *srpc;
1418 struct srpc_buffer *buffer;
1419 struct srpc_service *sv;
1420 struct srpc_msg *msg;
1421 enum srpc_msg_type type;
1423 LASSERT(!in_interrupt());
1428 spin_lock(&srpc_data.rpc_glock);
1429 if (ev->status != -ECANCELED) /* cancellation is not error */
1430 srpc_data.rpc_counters.errors++;
1431 errors = srpc_data.rpc_counters.errors;
1432 spin_unlock(&srpc_data.rpc_glock);
1434 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1435 ev->status, ev->type, errors);
1438 rpcev->ev_lnet = ev->type;
1440 switch (rpcev->ev_type) {
1442 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1443 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1445 case SRPC_REQUEST_SENT:
1446 if (!ev->status && ev->type != LNET_EVENT_UNLINK) {
1447 spin_lock(&srpc_data.rpc_glock);
1448 srpc_data.rpc_counters.rpcs_sent++;
1449 spin_unlock(&srpc_data.rpc_glock);
1451 case SRPC_REPLY_RCVD:
1452 case SRPC_BULK_REQ_RCVD:
1453 crpc = rpcev->ev_data;
1455 if (rpcev != &crpc->crpc_reqstev &&
1456 rpcev != &crpc->crpc_replyev &&
1457 rpcev != &crpc->crpc_bulkev) {
1458 CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1459 rpcev, crpc, &crpc->crpc_reqstev,
1460 &crpc->crpc_replyev, &crpc->crpc_bulkev);
1461 CERROR("Bad event: status %d, type %d, lnet %d\n",
1462 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1466 spin_lock(&crpc->crpc_lock);
1468 LASSERT(!rpcev->ev_fired);
1469 rpcev->ev_fired = 1;
1470 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1471 -EINTR : ev->status;
1472 swi_schedule_workitem(&crpc->crpc_wi);
1474 spin_unlock(&crpc->crpc_lock);
1477 case SRPC_REQUEST_RCVD:
1478 scd = rpcev->ev_data;
1481 LASSERT(rpcev == &scd->scd_ev);
1483 spin_lock(&scd->scd_lock);
1485 LASSERT(ev->unlinked);
1486 LASSERT(ev->type == LNET_EVENT_PUT ||
1487 ev->type == LNET_EVENT_UNLINK);
1488 LASSERT(ev->type != LNET_EVENT_UNLINK ||
1489 sv->sv_shuttingdown);
1491 buffer = container_of(ev->md.start, struct srpc_buffer, buf_msg);
1492 buffer->buf_peer = ev->initiator;
1493 buffer->buf_self = ev->target.nid;
1495 LASSERT(scd->scd_buf_nposted > 0);
1496 scd->scd_buf_nposted--;
1498 if (sv->sv_shuttingdown) {
1500 * Leave buffer on scd->scd_buf_nposted since
1501 * srpc_finish_service needs to traverse it.
1503 spin_unlock(&scd->scd_lock);
1507 if (scd->scd_buf_err_stamp &&
1508 scd->scd_buf_err_stamp < ktime_get_real_seconds()) {
1509 /* re-enable adding buffer */
1510 scd->scd_buf_err_stamp = 0;
1511 scd->scd_buf_err = 0;
1514 if (!scd->scd_buf_err && /* adding buffer is enabled */
1515 !scd->scd_buf_adjust &&
1516 scd->scd_buf_nposted < scd->scd_buf_low) {
1517 scd->scd_buf_adjust = max(scd->scd_buf_total / 2,
1519 swi_schedule_workitem(&scd->scd_buf_wi);
1522 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1523 msg = &buffer->buf_msg;
1524 type = srpc_service2request(sv->sv_id);
1526 if (ev->status || ev->mlength != sizeof(*msg) ||
1527 (msg->msg_type != type &&
1528 msg->msg_type != __swab32(type)) ||
1529 (msg->msg_magic != SRPC_MSG_MAGIC &&
1530 msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1531 CERROR("Dropping RPC (%s) from %s: status %d mlength %d type %u magic %u.\n",
1532 sv->sv_name, libcfs_id2str(ev->initiator),
1533 ev->status, ev->mlength,
1534 msg->msg_type, msg->msg_magic);
1537 * NB can't call srpc_service_recycle_buffer here since
1538 * it may call LNetM[DE]Attach. The invalid magic tells
1539 * srpc_handle_rpc to drop this RPC
1544 if (!list_empty(&scd->scd_rpc_free)) {
1545 srpc = list_entry(scd->scd_rpc_free.next,
1546 struct srpc_server_rpc,
1548 list_del(&srpc->srpc_list);
1550 srpc_init_server_rpc(srpc, scd, buffer);
1551 list_add_tail(&srpc->srpc_list,
1552 &scd->scd_rpc_active);
1553 swi_schedule_workitem(&srpc->srpc_wi);
1555 list_add_tail(&buffer->buf_list,
1556 &scd->scd_buf_blocked);
1559 spin_unlock(&scd->scd_lock);
1561 spin_lock(&srpc_data.rpc_glock);
1562 srpc_data.rpc_counters.rpcs_rcvd++;
1563 spin_unlock(&srpc_data.rpc_glock);
1566 case SRPC_BULK_GET_RPLD:
1567 LASSERT(ev->type == LNET_EVENT_SEND ||
1568 ev->type == LNET_EVENT_REPLY ||
1569 ev->type == LNET_EVENT_UNLINK);
1572 break; /* wait for final event */
1574 case SRPC_BULK_PUT_SENT:
1575 if (!ev->status && ev->type != LNET_EVENT_UNLINK) {
1576 spin_lock(&srpc_data.rpc_glock);
1578 if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1579 srpc_data.rpc_counters.bulk_get += ev->mlength;
1581 srpc_data.rpc_counters.bulk_put += ev->mlength;
1583 spin_unlock(&srpc_data.rpc_glock);
1585 case SRPC_REPLY_SENT:
1586 srpc = rpcev->ev_data;
1587 scd = srpc->srpc_scd;
1589 LASSERT(rpcev == &srpc->srpc_ev);
1591 spin_lock(&scd->scd_lock);
1593 rpcev->ev_fired = 1;
1594 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1595 -EINTR : ev->status;
1596 swi_schedule_workitem(&srpc->srpc_wi);
1598 spin_unlock(&scd->scd_lock);
1608 memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1609 spin_lock_init(&srpc_data.rpc_glock);
1611 /* 1 second pause to avoid timestamp reuse */
1612 set_current_state(TASK_UNINTERRUPTIBLE);
1613 schedule_timeout(cfs_time_seconds(1));
1614 srpc_data.rpc_matchbits = ((__u64)ktime_get_real_seconds()) << 48;
1616 srpc_data.rpc_state = SRPC_STATE_NONE;
1618 rc = LNetNIInit(LNET_PID_LUSTRE);
1620 CERROR("LNetNIInit() has failed: %d\n", rc);
1624 srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1626 LNetInvalidateEQHandle(&srpc_data.rpc_lnet_eq);
1627 rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1629 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1633 rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1635 rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1638 srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1646 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1658 state = srpc_data.rpc_state;
1659 srpc_data.rpc_state = SRPC_STATE_STOPPING;
1664 case SRPC_STATE_RUNNING:
1665 spin_lock(&srpc_data.rpc_glock);
1667 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1668 struct srpc_service *sv = srpc_data.rpc_services[i];
1670 LASSERTF(!sv, "service not empty: id %d, name %s\n",
1674 spin_unlock(&srpc_data.rpc_glock);
1678 case SRPC_STATE_EQ_INIT:
1679 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1680 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1682 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1683 LASSERT(!rc); /* the EQ should have no user by now */
1685 case SRPC_STATE_NI_INIT: